CONFIDENTIALITY-FIRST ARCHITECTURE DEEP DIVES

Verified Enterprise & AI Systems Proof of Work

Every architecture below pairs owned AI platforms and synthetic demonstrations with anonymized enterprise engagements, detailing warehouse schemas, incremental dbt watermarks, semantic layer bridges, and multi-stage Gemini pipelines.

CASE STUDY 01 // PERSONAL AI & APPLIED ANALYTICS // MULTIMODAL AI & BIGQUERY ML

AI Nutrition & Ingredient Intelligence: Built From Care, Guided by Evidence

⚡ Personal AI & Analytics Project
4 Modalities
Telegram + React 19 Capture
410 / 410
Zod-to-SQLite Synthetic Parity
24h / 48h
Prior-Only dbt & BQML Windows

Architectural Challenge & Business Bottleneck

Built to help my wife better understand how meals, ingredient quality, preparation, sleep, physical activity, and daily routines relate to how she feels. Everyday observations rarely arrive as clean rows, making weeks of overlapping habits hard to summarize for healthcare conversations.

Production Stack

React 19FirebaseTelegram Bot APIGoogle GeminiFatSecret APIdbt CoreBigQuery MLSemantic Scholar APIZod

Engineered Solution & Data Pipeline Topology

  • ▹Built multimodal conversational capture across Telegram (text, voice memos, plate photos, barcodes) and React 19 with a 3-stage Gemini Gatekeeper that asks one concise follow-up when missing prep or timing context matters (0 partial rows written in tests).
  • ▹Paired FatSecret OAuth 2.0 lookups and deterministic FDA/Monash additive rules (polyols, gums, emulsifiers) with shared Zod contracts and tenant-partitioned dbt Bronze -> Silver -> Gold 1 PRECEDING lag windows (410/410 synthetic SQLite parity).
  • ▹Combined exploratory BigQuery ML (XGBoost and ARIMA_PLUS) chronological holdouts, modeled baseline intervals, abs(r) >= 0.35 collinearity warnings, and Semantic Scholar MeSH literature references to surface honest questions for clinician review.
Verified Execution Flow:
Telegram + React 19 Intake→Gemini Gatekeeper + FatSecret Rules→dbt Bronze -> Silver -> Gold→BigQuery ML + Clinician Review Summary
CASE STUDY 02 // ENTERPRISE DATA PLATFORM & POWER BI SEMANTIC MODELING

Unifying B2B Healthcare Practitioner (HCP) Revenue & DTC Patient Demand Generation in BigQuery, dbt & Power BI

🔒 Publicly Traded B2B2C Biotech & Aesthetics Manufacturer
99.9%
Daily Partitioned Mart SLA
Zero Fanout
DAX V2 / TMDL Bridge Architecture
3-Day Lookback
Dynamic Watermark insert_overwrite

Architectural Challenge & Business Bottleneck

Commercial leadership needed a single source of truth separating B2B clinic/practitioner purchasing (hcp.* domains & loyalty tiers) from DTC consumer demand-generation media (Meta, Google Ads, GA4clinic finder flows), while preventing many-to-many fanout bugs in executive Power BI dashboards.

Production Stack

BigQuerydbtPower BI (TMDL / DAX V2)Deneb Vega-LiteGA4GTMPython

Engineered Solution & Data Pipeline Topology

  • ▹Architected a multi-layered BigQuery & dbt project (staging → intermediate → marts) with strict grain enforcement (campaign-level vs. brand-level vs. GA4 event-level).
  • ▹Engineered dynamic table watermark lookbacks (MAX(date) - 3) using BigQuery incremental insert_overwrite partitioning to handle late-arriving ad platform conversions at minimal slot cost.
  • ▹Designed the Power BI DAX V2 Semantic Model (.SemanticModel TMDL definitions) with isolated Strategy_KPI_Mapping calculation bridges and custom Deneb (Vega-Lite) executive KPI cards.
Verified Execution Flow:
GA4 + Paid Media APIs + Loyalty CRM→BigQuery Raw & Staging Views→dbt Incremental Marts (insert_overwrite, MAX(date)-3)→Power BI TMDL Semantic Model (DAX V2 + Deneb Visuals)
CASE STUDY 03 // EXPERIMENTATION GOVERNANCE & CONVERSION ARCHITECTURE

Cross-Platform Experimentation Telemetry, Read-Only Audit Automation & Statistical Decision Pipelines

🔒 Top-3 US Consumer Financial Wellness & Debt Relief Platform
100%
Experiment Schema Audit Coverage
Sub-5 Min
Automated Variant & Targeting Extraction
Full-Funnel
Lead-to-Enrollment Metric Alignment

Architectural Challenge & Business Bottleneck

High-velocity acquisition funnels running concurrent client-side and feature-flag experiments across AB Tasty and Statsig suffered from fragmented targeting rules, manual QA bottlenecks, and disconnected downstream enrollment attribution.

Production Stack

StatsigAB TastySegmentBigQueryDagsterPythonAgentic QA Skills

Engineered Solution & Data Pipeline Topology

  • ▹Built deterministic read-only API inspection and governance skills for AB Tasty and Statsig to automatically extract active test variations, JS/CSS code modifications, URL targeting scopes, and audience triggers.
  • ▹Standardized Cloud metric definitions, experiment hypothesis documentation, and automated executive decision synthesis connecting front-end variant exposure to qualified backend financial consultations.
  • ▹Integrated multi-channel communication triage and technical project management playbooks that convert meeting transcripts directly into sprint-ready engineering specifications.
Verified Execution Flow:
AB Tasty & Statsig Console APIs→Automated Read-Only Governance & Targeting Auditor→Segment / Warehouse Exposure & Conversion Join→Executive Experiment Readout & Rollout Gate
CASE STUDY 04 // D2C TELECOM ANALYTICS ENGINEERING & EXPERIMENTATION

Automated Product & CRO Tracking Architecture (-50% Manual Tagging) + Offline Call-Center LTV/CAC Unification

🔒 National D2C Wireless Carrier (Mint Mobile / Telecom Scale)
-50%
Engineering Manual Tagging Effort
360° View
Web + App + Call Center Attribution
+40%
Faster GTM & Experiment Launch Cycles

Architectural Challenge & Business Bottleneck

Rapid subscriber growth required unifying high-volume GA360/GA4 web and mobile app telemetry with offline call-center conversions (Invoca) while eliminating the 6-month engineering bottleneck caused by manual element tagging for product analytics and A/B tests.

Production Stack

SnowflakeBigQuerydbtLookerAmplitudeOptimizelyServer-Side GTMInvoca

Engineered Solution & Data Pipeline Topology

  • ▹Led a cross-functional initiative with Engineering and Product to embed an automated site tracking wrapper (Ampli / server-side GTM) that enabled product analytics and Optimizely/Amplitude experimentation without manual element tagging-cutting manual effort by 50%.
  • ▹Owned the GA360/GA4-to-BigQuery and Snowflake dbt modeling strategy, flattening nested event structures and joining offline call-center conversions to map the full subscriber lifecycle from lead to renewal.
  • ▹Built executive Looker dashboards and cohort LTV/CAC models across segmented B2C subscriber plans to guide C-suite (CPO/CRO) budget allocation.
Verified Execution Flow:
Web/App Ampli Wrapper + sGTM + Invoca Call Center→BigQuery & Snowflake Raw Event Streams→dbt Lifecycle, Churn & LTV:CAC Cohort Models→Looker Executive Dashboards + Amplitude / Optimizely CRO
⚡ Internal AI Architecture Lab
50 States
3-Tier Census Community Impact Cache
Gemini 3.x
Flash-Lite / Flash / Pro Model Router
1-Time
Cross-Bill Cumulative Voter Profile

Architectural Challenge & Business Bottleneck

Full congressional bills run hundreds of pages of dense legal text. Citizens need an unbiased, personalized breakdown of how a bill impacts their household and community (ZIP/City/State), plus an NPS-style alignment score comparing their section-by-section votes against their Representative and Senators-without incurring runaway LLM compute costs.

Production Stack

Gemini 3.8 Flash / 3.1 ProBigQueryFirestoreFastAPI (Cloud Run)Next.js 16TypeScriptUS Census ACS

Engineered Solution & Data Pipeline Topology

  • ▹Engineered a Two-Tier Lakehouse-to-Storefront architecture: BigQuery (bill-analyzer-3ccad.discover_vote) stores legislative roll-call votes, legislator profiles, BLS/EIA economic indicators, and US Census ACS 5-Year demographics, syncing incrementally via watermark timestamps to Firestore.
  • ▹Built OntologyLearner using Stage 1 fuzzy token matching (SequenceMatcher >= 0.70) + Stage 2 Gemini semantic verification to deduplicate demographic prompts across bills, paired with a cumulative localStorage voter profile (DemographicForm.tsx) that asks each question only once and auto-populates across future bills.
  • ▹Implemented 3-Tier Hierarchical Community Impact Caching ({bill}_{zip} → {bill}_{city_state} → {bill}_state_{state}) with asynchronous granular ZIP upgrade and bulk 50-state pre-fetching.
Verified Execution Flow:
Congressional Bills + Roll Call Votes + US Census ACS (BigQuery)→BillProcessor + OntologyLearner (Gemini 3.1 Flash-Lite)→Incremental Watermark Sync → Firestore & GCS Storefront→Next.js 16 Civic Portal + 3-Tier Cached Gemini 3.8 Impact Engine